Google Ships Gemini 4 Carbon and a Swarm of Coding Agents

Google made Gemini 4 Carbon, which it calls its strongest model, available to developers, businesses, scientists, and consumers on November 10, 2026. If you use the Gemini app, the catch is the plan. Carbon is rolling out to Google AI Ultra subscribers in the Gemini app and in AI Mode in Search, and Ultra starts at $99.99 a month in the US. Google AI Pro, the $19.99 plan, is not named in the launch post, and Google’s plans page lists it as including Argon, not Carbon.

If you build software, the catch is a date. Carbon costs $2 per million input tokens and $10 per million output tokens through November 30, 2026, then $4 and $20 starting December 1. That is the same introductory price as Gemini 4 Argon, the sibling model Google released about a month ago.

The other half of the launch is a feature called /swarm, rolling out in preview in Google Antigravity, the company’s coding tool. Type /swarm and one prompt goes to a team of artificial intelligence (AI) agents: a coordinator, workers that build in parallel, and verifiers that test the result.

Google is not telling everyone to switch. Its launch post says Argon “remains the ideal choice for the majority of software engineering and knowledge work,” while Carbon is built for the hardest jobs and uses more tokens to get there.

Who Gets Gemini 4 Carbon at Launch

Ultra comes in two sizes in the US: $99.99 a month with five times the Gemini usage limits of the Pro plan, and $199.99 a month with 20 times. Google’s plans page says Carbon is available to Ultra subscribers “in select countries,” and that only personal Google Accounts can sign up for its AI plans.

Paying for Ultra does not make the Gemini app unlimited. Google’s help page on app limits, which does not name Carbon, says those limits are based on computing use, refresh every five hours until you reach a weekly limit, and count the complexity of your prompt, the features you use, and the length of your chat. The same page says raising a model’s effort level in the app gets more thorough answers, but it “will also use more of your limit.”

The science offer builds on an earlier one. Four weeks before this launch, Google opened applications to give 10,000 scientists free access to Argon in Antigravity for six months, and it is now adding Carbon for current and future participants. Security teams apply separately to the Fairwind Program, which is how a first group of cyber defenders got Carbon two weeks before everyone else.

What It Costs Developers

Carbon is not on the free tier of the Gemini Application Programming Interface (API). Google’s pricing page adds that “every token the model processes is counted and billed,” including the thinking tokens the model spends before it answers.

That last part is where Carbon and Argon differ on your bill. The two share a per-token price, and Google’s developer guide says the main difference is how many thinking tokens each spends per turn: Carbon spends more, “which increases per-turn token use and latency compared to Argon.” You control some of that with a thinking level of low, medium, or high, and medium is the default. Google’s own advice is to “run your evaluations at all three levels, then pick the lowest level that meets your quality bar.”

One independent measurement shows what Carbon costs per task. Artificial Analysis, a company that tests AI models, reports that running its Intelligence Index cost $7.76 per task on Carbon at high thinking, $6.59 at medium, and $7.19 at low. Its page lists Carbon at the $4 and $20 rate.

How /swarm Works

Google’s Antigravity team describes /swarm as a fix for a bottleneck. With earlier subagents, one main agent handed out tasks and then had to review and route every step itself. In a swarm, a coordinator starts specialized workers “in isolated git worktrees so edits never collide,” and separate verifiers run tests and challenge the changes. The agents talk to each other on a shared message board and wait at “zero token cost when idle.”

You can watch the whole thing. A Swarm panel in Antigravity shows a live Board feed of the agents’ discussion, a Members roster where you can open any agent’s work and steer it, and a Graph timeline that replays how the team got from the brief to tested code. In the command line tool, /swarm board or Alt+S opens a full-screen version.

Each agent in a swarm can run on a different model. Google suggests putting Carbon in the coordinator, architecture, security auditing, and verification roles, while “lighter, faster models handle the parallel file edits and routine implementation tasks.”

Two limits are worth knowing before you try it. /swarm is a preview, rolling out in the Antigravity App and command line tool. And Google attaches what it calls a fair warning: parallel multi-agent systems “consume quota much faster than a standard chat.”

Google’s examples are showpieces. It says a multi-agent Carbon team using /swarm turned seven public city council planning PDFs into an interactive 3D reconstruction of a Google office, and that another swarm built a browser game in which you explore Google’s Mountain View campus as a pelican riding a bike. Inside Google, engineers ran swarms for five weeks on jobs such as multi-package migrations and architectural redesigns.

What the Tests Show, and Who Ran Them

Google’s text does not give one clean table of scores, so sort its claims by who did the testing.

An independent index. Artificial Analysis scores Carbon at high thinking at 60 on its Intelligence Index, behind Claude Fable 5.2 at maximum effort at 65 and one point ahead of Claude Opus 5.5 at maximum effort at 59. Its summary ranks Carbon fourth of 74 models. Argon scores 50 on the same index. When Google first announced Carbon on October 28, it said the model ranked first on version 5.0 of that index with a score of 59. The index is now on version 5.0.1 and the chart has moved, so that first-place claim no longer matches what Artificial Analysis shows. This site covered Claude Fable 5.2 separately.

Rankings Google reports from other companies’ tests. Google says Carbon ranks first on CursorBench 4.0, Perplexity’s WANDR, and Zapier’s AutomationBench “at more than 2X lower cost per task compared to the next best external model.” The scores behind those rankings are not in the launch post.

A test Google helped build. The Multimodal Knowledge Agent benchmark, which checks how agents handle files full of charts, blueprints, and tables, was developed by Google with Artificial Analysis. Google says Carbon places second among the models tested, and that it finishes tasks at six times lower cost on average than the top scorer. Artificial Analysis runs the 200-task test set and keeps it private.

Security scores measured without guardrails. Google reports 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploiting them, and 71.0% on CWE-bench for patching them. The October post says those were “evaluated without guardrails,” and the launch post says Carbon leads on them “when accessed through the Fairwind Program.” On the patching test, Argon scored 68.0%.

Customer quotes. The launch post carries testimonials from companies including Hebbia, Glean, Box, Cisco, and Fortinet. They are the customers’ own accounts: HubX, for example, says Carbon resolved 94% of 16 real bug fixes in its codebase with zero regressions.

Google’s own research claims. Google says its researchers used Carbon alongside other models to turn security goals into theorems, then had agents built on a framework called AlphaProof Code use Carbon to write millions of lines of code in Lean, a language for mathematical proofs, proving key security properties of the code behind Signal’s encryption protocol in weeks. It describes that as “a task that historically takes expert teams years.”

What the Safety Report Says

Google published a model card and a separate Safety and Security Report for Carbon. Four points matter if you plan to rely on the model.

The public model has tighter security limits. The report says that, as with Argon, Google keeps “a conservative safety margin” on security requests, with extra refusals concentrated in “borderline dual-use reverse-engineering and exploit-adjacent auditing prompts rather than general software engineering workflows.” Carbon also “refuses slightly more frequently than Argon” on text prompts in Google’s over-refusal tests, though it is less likely to over-refuse on image prompts. Verified defenders in the Fairwind Program get mitigations tailored to allow tasks such as checking whether a vulnerability can be exploited.

Google reports strong numbers against attacks, with a caveat. In simulated cyberattack tests, Google says Carbon’s safeguards blocked 99% of attack paths from an automated red-teaming agent and cut the per-attempt breach rate to 0.8%, from 3.1% on Argon. On Gray Swan’s test of hidden instructions planted in content an agent reads, Carbon scored a 0% attack success rate. The same report says such attacks “on Carbon are still possible.”

It suspects it is being tested far more often. In one set of Google’s alignment tests, Carbon said out loud that it might be in a safety evaluation in 28.3% of runs, up from 1.7% for Argon. A model that knows it is being tested is harder to judge from test results, and the report says of the same tests that “unspoken evaluation awareness cannot be ruled out.”

Google did not run a full new risk assessment. Under its Frontier Safety Framework, Google found Carbon “did not show meaningful new capabilities or material increases in performance compared with Gemini 4 Argon,” so it carried forward Argon’s risk findings. The model card says Carbon is based on Argon and has a knowledge cutoff of March 2026.

What Changes for Developers

The model ID is gemini-4-carbon, with a 1-million-token window shared by input, thinking, and output. If you already run Argon, the switch “requires no request schema changes,” though Google still says to check that your output limit and timeouts allow for longer reasoning. Coming from Gemini 3 or earlier takes more work:

• Explicit caching is gone. Carbon does not support it, so calls that create or reference cached content have to come out; implicit caching is on by default for prompts of at least 4,096 tokens.
• Sampling settings are ignored. Google says to remove temperature, top_p, and top_k.
• Long answers need background mode. Standard responses stop at 65,536 tokens, and for longer ones Google recommends background=True with a higher max_output_tokens.
• Raise your limits and timeouts. Google recommends a max_output_tokens of at least 32,768, because Carbon can spend more of it on thinking, and either background execution or client timeouts of at least 15 minutes.

The model card lists occasional slowness or timeouts among Carbon’s known limitations. Artificial Analysis lists 67 seconds to the first answer token in its summary of the model.

What Comes Next

• November 30, 2026: the last day of the $2 and $10 introductory API price.
• December 1, 2026: the API price becomes $4 per million input tokens and $20 per million output tokens, for Carbon and Argon alike.
• No date given: /swarm is in preview, and Google’s posts do not say when it leaves preview.

Who Should Use It

If you pay for Google AI Ultra on a personal account, Carbon is rolling out to your plan in select countries, so once it appears, try it on a problem that the everyday models fumble and watch your usage as you go. If you are on Google AI Pro or no plan, Carbon is not in the Gemini app or Search for you at launch; the routes left are the paid API or, for researchers, the science program.

If you are a developer, the three weeks before December 1 are the cheap window to test Carbon against Argon on your own workload. Start at medium thinking, which is the default, test low and high too, and keep the lowest level that meets your bar. And if you want to try /swarm, do it on a contained project first: it is a preview, and Google itself warns that it eats quota quickly.

Sources (14)

• blog.google Google: Gemini 4 Carbon is now available
• blog.google Google: Gemini 4 Carbon, our next frontier model
• blog.google Google: Gemini 4 Carbon is now available for developers
• antigravity.google Google Antigravity: Gemini 4 Carbon and /swarm
• blog.google Google: Multimodal Knowledge Agents benchmark
• ai.google.dev Gemini API: Pricing
• ai.google.dev Gemini API: Models
• ai.google.dev Gemini API: Release notes
• one.google.com Google One: Google AI plans
• artificialanalysis.ai Gemini 4 Carbon
• storage.googleapis.com Google DeepMind: Gemini 4 Carbon Model Card
• storage.googleapis.com Google DeepMind: Gemini 4 Carbon Safety and Security Report
• aistudio.google.com Gemini API: What's new in Gemini 4 Carbon
• support.google.com Google Help: Changes to Gemini model access and limits